Segmenting anatomical structures is an important part of the analysis of image data, in particular of medical image data as well. Important examples are the planning of surgical procedures, volumetric examinations of organs, evaluating the progression of metastases, or the statistical modeling of organs.
In this case, image information about a three-dimensional (3D) structure is usually present as a sequence of two-dimensional (2D) scan slices of an imaging modality, such as, for example, computed tomography (CT) or magnetic resonance imaging (MRI). For this reason, it is conventional that the desired structure must firstly be segmented in the individual scan slices before it can be composed to form a 3D structure. However, segmentation algorithms for directly segmenting 3D structures are already available as well.
A fully automatic segmentation of medical image data is barely possible to implement using current technology. For this reason it is inevitable that a user must be able to independently intervene during the segmentation. Depending on the segmentation algorithm used, the user intervenes in different ways in the segmentation in order to control the result in the desired manner.
Diverse algorithms for interactive segmentation exist. Here, inter alia, two different main groups can be distinguished: contour-based segmentation algorithms and region-based segmentation algorithms.
As an exemplary contour-based segmentation algorithm, the so-called livewire method will be discussed briefly here. This method has become established as a qualitatively high-grade interactive segmentation of organ structures from 3D data records. It is very reliable, particularly in the case of high image contrasts. However, in the case of weak image contrasts an extremely large number of interventions by the user is often required to achieve an acceptable result.
The basic concept of the livewire method is the following: a user marks a starting point on the contour of the structure in a displayed image data record, for example by way of a cursor and a mouse, and then moves the cursor to another position in the image data record. The livewire algorithm calculates the course of the contour from the starting point to the current position of the cursor. For this purpose, a so-called cost function is used which allows a path to be calculated which satisfies certain criteria in an optimum manner, such as, for example, minimum change of the value of the gradient along the path. Should the calculated path not lie correctly on the contour of the structure, the user can take corrective action, for example by simply clicking on and displacing the path. More details are described, for example, in W. A. Barret and E. N. Mortensen (1997) “Interactive Livewire Boundary Extraction”, Medical Image Analysis, Vol. 1, No. 4, pp. 331341, CVPR '99 #107 page 6/7.
As an example of a region-based segmentation algorithm, the so-called GraphCut method will be discussed briefly here. This method too has become established as qualitatively high-grade and achieves a good result even in the case of weak image contrasts. In the GraphCut method, the user characterizes such image regions which are located within the structure and such image regions which are located outside of the structure. The GraphCut algorithm calculates the maximum discontinuity between these regions, again using a cost function as a criterion which comprises, for example, gray scale information in the characterized regions. This maximum discontinuity corresponds to the border of the structure. If the result is not yet satisfactory after a first segmentation, further inner and outer regions can be marked until an acceptable segmentation result is present. A more detailed description of a GraphCut algorithm and an associated cost function are disclosed in US 2004/0008886 A1, for example.
By way of such segmentation algorithms, the user works through a given 3D image data record slice by slice until the entire structure is segmented. Depending on the segmentation algorithm used and the image contrasts present in each case, the user can often be forced to intervene, which can significantly increase the processing time of a 3D image data record.
Therefore, there is still the need for user-friendly segmentation algorithms which allow rapid segmentation of structures in a manner which is as intuitive as possible and requires little interaction.